Google’s already investing heavily in machine learning, a type of artificial intelligence in which computers, given a large enough mass of data to chew through, can detect patterns on their own and make judgements about how to handle the information. Machine learning powers Gmail’s spam filters and the Google Now service, which tries to predict what information a user will find helpful or interesting.
On Tuesday, Google announced that machine learning will be put to use in a new way: automatically responding to e-mails for you.
The “Smart Reply” feature will roll out later this week to Inbox, Google’s slightly-experimental e-mail service focused on organization and time management. Smart Reply will automatically parse the text of received e-mails and suggest three responses that can be inserted with a click or a tap.
“For those emails that only need a quick response, it can take care of the thinking and save precious time spent typing,” software engineer Bálint Mikló wrote in a blog post. “And for those emails that require a bit more thought, it gives you a jump start so you can respond right away.”
There’s a lot going on under the hood of Smart Reply. At the core of the feature is a pair of deep neural networks, which are sets of algorithms that processes data at several layers at once to learn sentence structure, writing style, and tone.
“These systems generalize better, and handle completely new inputs more gracefully than brittle, rule-based systems ever could,” Google senior research scientist Greg Corrado wrote in a separate blog post.
One neural network parses the text of the incoming e-mail to capture a list of “thought vectors” – essentially, the gist of what’s being said or asked. If the e-mail contains a question, the second network will compose a simple reply that provides information to answer the query. If an e-mail asks, for example, “Do you have vacation plans yet,” Smart Reply might suggest as responses, “No plans yet,” “I just sent them to you,” and “I’m working on them.” The system also judges the intent of the proposed responses so that it doesn’t generate overlapping options that use different phrasing to say the same thing.
Smart Reply also gets more accurate with use, since the networks learn from larger bodies of text. Google’s early prototype had the propensity to respond to buttoned-down business e-mails with “I love you," since the system had identified that as a common phrase and would fall back on it when it was unsure of how to respond. The team fixed that issue by having Smart Reply check to see how often phrases had been used previously.